A neural network model for simulation of water levels at the Sultan Marshes wetland in Turkey


Dadaser-Celik F., Cengiz E.

WETLANDS ECOLOGY AND MANAGEMENT, vol.21, no.5, pp.297-306, 2013 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 21 Issue: 5
  • Publication Date: 2013
  • Doi Number: 10.1007/s11273-013-9301-y
  • Journal Name: WETLANDS ECOLOGY AND MANAGEMENT
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus
  • Page Numbers: pp.297-306
  • Keywords: Wetlands, Water levels, Modeling, Artificial neural networks, Sultan Marshes, Turkey, SWAT
  • Erciyes University Affiliated: Yes

Abstract

An artificial neural network (ANN) model was developed for simulating water levels at the Sultan Marshes in Turkey. Sultan Marshes is a closed basin wetland located in the semi-arid Central Anatolia region of Turkey. It is one of the thirteen Ramsar sites of Turkey and a national park. Water levels at the Sultan Marshes showed strong fluctuations in recent decades due to the changes in climatic and hydrologic conditions. In this study, monthly average water levels were simulated using a multi-layer perceptron type ANN model. The model inputs consisted of climatic data (precipitation, air temperature, evapotranspiration) and hydrologic data (ground water levels, spring flow rates, and previous month water levels) available from 1993 to 2002. 70 % of the data were used for model training and remaining 30 % were used for model testing. Model training was accomplished by using a scaled conjugate gradient backpropagation algorithm. The performance of the model was evaluated by calculating the root mean square error (RMSE) and the coefficient of determination (R (2)) between observed and simulated water levels. The sensitivity of the model to input parameters was determined by evaluating the model performance when a single input variable was excluded. It was found that the ANN model can successfully be used for simulating water levels at the Sultan Marshes. The model developed using all input variables provided the best results with two neurons in the hidden layer. The RMSE and R (2) of the simulated water levels were 4.0 cm and 96 %, respectively. The sensitivity analysis showed that the model was most sensitive to previous month water levels and ground water levels.